| The photovoltaic(PV)power generation is affected by the light condition which varies according to the weather.That is the reason why PV power generation is highly random and fluctuating.Nowadays,the use of PV power generation is increasing with the advance of the power grid.For the safety and accuracy of the dispatch,the forecasting of PV power generation is greatly required.For the time series of photovoltaic data,the relationship between the BP neural network and the information in the time dimension is not added,resulting in unpredictable prediction accuracy.RNN(recursive neural network)structures like LSTM and GRU are widely used in time series prediction.RNN has a nested structure and has a good prediction effect,but lacks the ability of parallel computing.Inputs at all times must traverse the network and pass from beginning to end.Facing the problem of long training time,the sequence will be forgotten if the transmission is too long in the network.Important timing information.This paper proposes a method of photovoltaic prediction using SRNN sliced neural network to ensure the ability of parallel computing by segmenting the RNN network without affecting the sequence information,and because RNN is cut,the number of effective information transmission in the network is reduced and it is not easy to forget,So that the accuracy of the model is improved.This article first describes and analyzes some of the photovoltaic prediction literature,introduces the background and significance of photovoltaic prediction,and the problems of the commonly used neural network method,and then introduces the principle of the method used in the model.Through modeling and comparison verification,the results It shows that the SRNN network is much faster and more accurate than the LSTM and GRU common models in photovoltaic prediction,which proves the superiority of the SRNN prediction model.The specific content is.(1)Establishment of the data set.The data from a power station in Northwest China for 18 years including missing data filling,iforest outlicr data removal and data split is pre-processed.Then,the data is used as the input to the network for the establishment of data sets.(2)Comparison of experimental simulation data for the establishment of forecasting model.Firstly,the structure of the network is determined as three-layer GRU plus FC fully connected laycr.Then the capacity of GRU and FC was determined through multiple training.L2 regularization is then added to prevent the model from improving its accuracy.By the compared experiment of the optimizer,we select the optimizer Ada,to train the network.(3)Comparison with other commonly used deep learning mcthods.It proves that SRNN has huge advantages in both training speed and convergence speed.Due to the facts that RNN has fast training and convergence speed,its model accuracy is also significantly higher than that of other deep learning methods,which proves the advantages of SRNN networks. |
